[CUDA] Register BF16 ReduceMean kernels (#32326)
## Description
Register the existing CUDA BF16 `ReduceMean` kernels with the CUDA
execution provider. The kernel implementations already cover both the
attribute-based and axes-input schema ranges, but their create-info
entries were missing from the provider registry.
## Summary of Changes
- Register BF16 `ReduceMean` for opsets 1 through 17.
- Register BF16 `ReduceMean` for opset 18 and later.
- Add CUDA-only regression tests for opsets 13 and 18 using an explicit
CUDA execution provider.
## Testing
- Compiled `reduction_ops_test.cc` with the CUDA Debug configuration and
`-Werror`.
- Executed BF16 `ReduceMean` models for opsets 13 and 18 with CPU
fallback disabled; both produced the expected output on CUDA.
## Checklist
- [x] Tests added/updated
- [x] No breaking changes
- [x] Documentation not required
